Update app.py

app.py

@@ -8,7 +8,12 @@ import os
 llm = HuggingFaceEndpoint(
     repo_id="HuggingFaceH4/zephyr-7b-beta",
     temperature=0.7,
-    model_kwargs={"max_length": 512},
+    model_kwargs={
+        "max_length": 256,  # Reduced from 512
+        "max_new_tokens": 256,  # Add explicit limit
+        "do_sample": True,
+        "top_p": 0.9
+    },
     huggingfacehub_api_token=os.getenv("HUGGINGFACEHUB_API_TOKEN")
 )